Groundwater Heights Prediction from Seismic Waves with Machine Learning

Abstract

Unlike surface water reservoirs, that can be easily quantified and monitored, underground conduits in karst systems are often inaccessible, hence challenging to monitor. Seismic noise analysis was proved to be a reliable tool to monitor ground water storage in a fractured rock aquifer (Lecocq et al. 2017). In underground karstic environments, seismic noise monitoring was able to detect hydrological cycles and monitor the groundwater-content variations (Almagro Vidal et al. 2021). The following approach relies on coupling passive seismic wavefield with hydrological data in a machine learning algorithm in order to monitor underground water heights. The studied site is the Fourbanne karst aquifer (Jura Mountains, Eastern France, Jurassic Karst observatory). The underground conduit is accessible through a drilled shaft and instrumented by two 3-component seismological stations, one located underground and the other one at the surface, and a water height probe. We applied a new approach based on the machine learning random forest (RF) algorithm and continuous seismic records (Hibert et al., 2017), to find characteristic signals to predict the underground river water height. The method consists on the computation on a sliding window of seismic signal features (waveform, spectral and spectrogram features) and using the corresponding water height at the same time window to train the algorithm, and then apply it on new data. The RF algorithm is capable of accurately detecting flooding periods and reproduce the groundwater heights with an efficiency exceeding 95% and 53% using the Nash-Sutcliffe criterion for the seismic stations located in the underground conduit and at the surface respectively. The obtained results are a first promising outcome for the remote study of water circulation in karst aquifers using seismic noise.